DaYa-YOLO: Substituting a Frozen Colorimetric Prior for Learned Attention in Edge-deployed Rice Pest and Disease Detection
Rice pest detection must run on affordable handheld hardware, yet color space and edge deployability have been studied separately.We ask whether a frozen color prior can replace learned attention.DaYa-YOLO pairs an RGB backbone with a frozen sRGB to CIELAB transform and a trainable scaling layer, compared against three attention modules over five seeds on 13 field-collected rice classes.DaYa-LAB reaches 0.707 mean mAP50 against the YOLO11 baseline's 0.709, and none of the eleven architectures separates on mean accuracy, though the dual-branch and single-branch architectures differ in neck wiring, so this is accuracy parity rather than controlled substitution.An architecture-matched control without the color transform reaches 0.705 and is equally indistinguishable, so the transform adds no measurable accuracy.Robustness falls for every dual-branch model.Deployment separates them: 8 of 10 produced a working Jetson Nano engine.Reproducible latency and compilation are practical deployment criteria.
Authors
- Refa Firgiyanto (ORCID: https://orcid.org/0000-0003-0883-9508)
- Gallyndra Fatkhu Dinata (ORCID: https://orcid.org/0000-0002-4968-324X)
- Mike Yuliana (ORCID: https://orcid.org/0000-0002-8751-560X)
- Nailul Muna
- Moch. Zen Samsono Hadi (ORCID: https://orcid.org/0000-0001-7732-9996)
- Hendri Darmawan
- Riyadh Hadinah Ahtar Lakadimu
- Dwi Putro Sarwo Setyohadi
- I Kadek Dipastra Arka Dananjaya
Institutions
- Politeknik Elektronika Negeri Surabaya (ID)
- Politeknik Negeri Jember (ID)
Publication Details
- Journal
- International journal of intelligent engineering and systems
- Published
- 2026-11-17
- DOI
- https://doi.org/10.22266/ijies2026.1231.02
- Primary Topic
- Smart Agriculture and AI
- Type
- article
- Field-Weighted Citation Impact
- 0.00